antirez-recipe quants
MoE quants following the antirez recipe: routed experts quantized hard, everything else protected.
| Tensor class | Quant | Notes |
|---|---|---|
blk.*.ffn_gate_exps, blk.*.ffn_up_exps (ffn_gate_up_exps if fused) |
GU_TYPE (imatrix) | routed-expert gate/up |
blk.*.ffn_down_exps |
DOWN_TYPE (imatrix) | routed-expert down (K-quant for quality) |
blk.*.ffn_{gate,up,down}_shexp, all attn_*, output.weight |
Q8_0 |
shared experts / attention / head |
token_embd.weight |
F16 |
embeddings |
router (ffn_gate_inp), norms, sinks, biases |
source precision (BF16/F32) |
never quantized by llama.cpp |
Base type Q8_0 + --tensor-type overrides, imatrix calibration
(see SPEC per-model imatrix links). Runs on upstream llama.cpp — no fork or
patch needed.
Validation per file: gguf-dump tensor audit (experts = GU/DOWN, protection intact),
llama-cli smoke load, llama-perplexity spot ppl + --kl-divergence vs BF16 logits.
Files
| File | Size | SHA-256 |
|---|---|---|
Ornith-1.5-35B-A3B-Q2-IQ2_XXS-Q2_K.gguf |
14215041696 | f6848c148c1d4644f103c0e1b735545c262f9779f83e9489b95235fadb345916 |
Imatrix: https://huggingface.co/mradermacher/Ornith-1.5-35B-A3B-i1-GGUF
Recipe
Antirez recipe on ornith-ai/Ornith-1.5-35B-A3B (BF16 source, ornith-ai/Ornith-1.5-35B-A3B):
ffn_gate_exps/ffn_up_exps/ffn_gate_up_exps-> IQ2_XXSffn_down_exps-> Q2_K- attention + shared experts + repr-MTP blocks blk.38/39/40 + output -> Q8_0
- token embeddings -> F16
- imatrix: mradermacher/Ornith-1.5-35B-A3B-i1-GGUF
Notes: mradermacher's imatrix does not cover the MTP/draft blocks (blk.38-40); they are forced to Q8_0 with a first-position llama-quantize rule (first matching regex wins). ~14.2 GB, about 2.9 BPW.
Validation (llama.cpp, qwen35moe arch)
- 12-case hard-smoke eval (temp 0, ctx 15360, model's own embedded chat template): 9/12 PASS; ~27 t/s generation (AMD Vulkan/UMA); fails cases 10567, 1606, juliet-cwe121. Passes both big-code cases.
- Exact output tokens (own tokenizer) over the 9 solved: 51,528 = 5,725 tok/solved.
- KL divergence vs official ornith-ai Q8_0 (ds4 short prompts, ctx 60, 2 chunks, CPU): per-chunk KL 0.214 / 0.262 (avg 0.238), delta-p RMS ~10.8%, same-token top-1 86.2 / 81.0%.